Select challenging examples from the dataset

Select the most relevant strategies for a given problem to be used during inference. This controls how many strategies are included in the sy

Set up performance test fixtures

Set the final synthesized solution

Set the steering strength for a specific pattern. Args: pattern: The reasoning pattern strength: The

Setup Chrome driver with appropriate options

Use tokenizer with its default configuration for inference

Share effective strategies across agents and generate enhanced solutions

Determine if we should continue with another iteration

Determine whether to create a new strategy or update an existing one. Args: problem_type: The type of problem query: The

Decide whether to execute code from the request based on the query.

Determine if current trace should be terminated based on confidence. Args: current_confidence: Current confidenc

Ask LLM to simulate code execution.

Suggest MLX alternative for a given model

Test that all approaches can be imported

Test that approaches handle parameters correctly

Approaches must degrade gracefully on empty / None / length-truncated provider responses instead of raising raw IndexError/TypeError/AttributeErro

Test basic approach calls with mock client

Test deep research plugin module

Test deepthink plugin and its submodules can be imported

Test genselect plugin module

Test real connection to GitHub MCP server via local stdio transport

Test longcepo plugin module

Test majority voting plugin module

Test that MARS uses different temperatures for agents

Test MCP plugin module

Test memory plugin has required structure

Test the n parameter with different values

Test that plugins load without relative import errors

Test plugin approach detection after loading

Test that plugin modules can be imported

Test all plugins with subdirectories can import their submodules

Test that multiple invocations of the privacy plugin don't have degraded performance. This catches the actual performance issue even without

Test that expensive resources (AnalyzerEngine, AnonymizerEngine) are created only once and reused across multiple plugin invocations.

Test proxy plugin module

Test that proxy plugin properly configures timeout settings

Test that proxy plugin handles timeouts correctly

Test that proxy plugin returns complete token usage information

Test that recognizers are not fetched/reloaded on each analyze() call. This prevents the performance regression where "Fetching all recognize

Direct test that singleton instances are the same object across calls.

Test spl plugin module

Test web search plugin module

Main plugin execution function with ThinkDeeper's controlled thinking process

MLX-compatible ThinkDeeper processing function

Convert the strategy to a dictionary for serialization.

Track request latency

Update component fitness based on performance (self-evolution) NOTE: This method tracks fitness values but they are not yet used to

Update the strategy text with a refined version.

Update the performance metrics for a strategy.

Validate if adapter exists and is compatible

Validate that citations are actually used in the report text. Checks which citation numbers appear in the body text and warns about unus

Verify if the solution contains the correct answer format for problems with specific answers

Check for problem-specific key insights that should appear in correct solutions

Verify a proof using IMO25-style two-stage verification Returns score on 0-7 scale and detailed assessment

Verify a solution using mathematical reasoning

Perform weighted majority voting based on trace confidences. Args: traces: List of trace results

Wrap reasoning in <think> tags and append clean final answer Args: reasoning: All intermediate reasoning, logs, agent outputs

Apply steering to the output of a layer. Args: module: The module being hooked input_tensors: The in

Process messages with AutoThink's controlled thinking.

Context manager entry - ensures browser is ready

Context manager exit - ensures browser cleanup